All pull requests

Pull requests

• Prepare tests to run over multiple backends.

  #11444·copybara-service[bot] opened Sep 29, 2026Bot·

  ·

• Add a test in attention when the KV cache grows.

  #11443·copybara-service[bot] opened Sep 29, 2026Bot·

  ·

• Extract buffer mapping logic that can be shared between XNNPACK and YNNPACK runners.

  #11442·copybara-service[bot] opened Sep 29, 2026Bot·

  ·

• Remove source_location shim now that absl::SourceLocation is open source.

  #11441·copybara-service[bot] opened Sep 29, 2026Bot·

  ·

• Automated Code Change

  #11439·copybara-service[bot] opened Sep 29, 2026Bot·

  ·

• Use 64-bit loads for the unpaired row in odd-MR NEON I8MM GEMM/IGEMM microkernels

  #11438·copybara-service[bot] opened Sep 29, 2026Bot·

  ·

• Add wasmrelaxedsimdfp16 and scalar fallback microkernels for F16 GEMM/IGEMM.

  #11436·copybara-service[bot] opened Sep 29, 2026Bot·

  ·
  1

• qu8 c8 weight packing for AVX2 and AVX256VNNI

  #11435·copybara-service[bot] opened Sep 28, 2026Bot·

  ·

• Check for null runtime in xnn_get_runtime_profiling_info and xnn_invoke_runtime.

  #11434·copybara-service[bot] opened Sep 28, 2026Bot·

  ·

• Zero-pad the last row slot in the QS8 scalar packw NC-remainder main K loop.

  #11430·copybara-service[bot] opened Sep 28, 2026Bot·

  ·

• Pack A in run_dot.

  #11429·copybara-service[bot] opened Sep 28, 2026Bot·

  ·

• Gate the sanitizer suppressions on the attribute, not on __has_feature

  #11426·anagnorisis2peripeteia opened Sep 28, 2026Contributor·

  ·

• Apply -ffp-contract=off to Clang builds, not just GNU

  #11425·anagnorisis2peripeteia opened Sep 28, 2026Contributor·

  ·

• [test] Scale vunary tolerance with the input as well as the output

  #11423·anagnorisis2peripeteia opened Sep 28, 2026Contributor·

  ·

• [reference] Avoid misaligned bias stores in the qs8-qb4w GEMM packers

  #11421·anagnorisis2peripeteia opened Sep 27, 2026Contributor·

  ·
  2

• [reference] Fix misaligned bias stores in the bf16 f32 packers

  #11420·anagnorisis2peripeteia opened Sep 27, 2026Contributor·

  ·

• [reference] Avoid misaligned float and int32 stores in qb4w packing

  #11419·anagnorisis2peripeteia opened Sep 27, 2026Contributor·

  ·

• [math] Avoid undefined behaviour in saturating_rounding_shift_left_s32

  #11418·anagnorisis2peripeteia opened Sep 27, 2026Contributor·

  ·
  1

• litert: guard SpaceToDepth/DepthToSpace/Split against zero divisor in arithmetic.h

  #11417·destro4evr-rgb opened Sep 27, 2026Contributor·

  ·

• litert: guard Concatenation against rank-mismatched subsequent inputs

  #11416·destro4evr-rgb opened Sep 27, 2026Contributor·

  ·

• [operators] Guard convolution-nhwc tiled output size overflow in reshape_igemm

  #11415·destro4evr-rgb opened Sep 27, 2026Contributor·

  ·

• [operators] Guard deconvolution-nhwc tiled output size overflow

  #11414·anagnorisis2peripeteia opened Sep 27, 2026Contributor·

  ·

• pack-lh: fix size_t overflow in size/offset functions on 32-bit platforms

  #11413·destro4evr-rgb opened Sep 26, 2026Contributor·

  ·

• packq: harden packed-LHS size helpers against integer overflow; reject SIZE_MAX at call sites

  #11412·VIRUSBABAA opened Sep 26, 2026·

  ·
  1

• subgraph: harden qs8->qc8 convert channelwise-quantization size arithmetic against integer overflow

  #11411·VIRUSBABAA opened Sep 26, 2026·

  ·
  1